Antibiotic discovery is an important task for combating the growing number of bacterial infections. While several approaches have been recently leveraged to accomplish this task, effectively using the large structural diversity of the chemical space for this purpose remains challenging. In particular, deep learning (DL) algorithms have showed great promise for identifying potential antibiotics from chemical libraries, but the lack of interpretability and explainability of commonly used black box models hinders the ability to derive chemical insights from predictions and models that could guide a more efficient exploration of the chemical space. In a recent work, James J. Collins and colleagues leveraged graph search algorithms to bring explainability to graph neural networks â a class of DL algorithms â in order to facilitate the identification of novel putative structural classes of antibiotics that are effective against pathogen S. aureus.
The authors hypothesized that model predictions of antibiotic activity and human cell cytotoxicity could be explained on the level of chemical substructures. More specifically, because graph neural networks make predictions based on the information contained in the atoms and bonds of molecules, compounds with high antibiotic prediction scores, for instance, could contain chemical substructures that would largely determine their scores. The explainability approach was then developed using a Monte Carlo tree search algorithm: the algorithm iteratively looks for the substructures in a search tree, in which the root of the tree is the original molecule with high predicted antibiotic activity, and each state in the tree is a subgraph derived from a sequence of bond or ring deletions. After developing models for human cell cytotoxicity and for antibiotic activity against S. aureus, the authors used their explainability approach to identify structurally new antibiotic classes in more than 12âmillion compounds. Notably, one new class exhibited high selectivity, overcame resistance, and possessed some favorable toxicological and chemical properties. The authors tested compounds from this class in vitro and in vivo, experimentally demonstrating their effectiveness against the studied pathogen. All in all, the work makes the marriage between drug discovery and DL potentially more effective by helping researchers to focus on more promising chemical spaces.
This is a preview of subscription content, access via your institution
